Opus 4.7 capped at 200K context instead of 1M (GitHub Copilot offers 1M, Zed does not)

[anilpai]

Reproduction steps

1. Start Zed with Opus 4.7 (either via built-in assistant with Anthropic key or Copilot extension).
2. Load a coding project, interact in the assistant/chat, and let the conversation grow.
3. Once the context reaches ~144K tokens, observe Zed asking to start a new context, while the session should support up to 1M tokens (as seen in GitHub Copilot's Opus 4.7 integration with 1M context windows).

Current vs. Expected behavior

Current behavior: Zed assistant prompts for a new context at ~144K tokens, reflecting a 200K hard cap (72% threshold trigger), instead of supporting Opus 4.7's full 1M window. This limitation is likely due to legacy token denominator logic for Anthropic models from Opus 4.6.

Expected behavior: Zed should allow the context window to reach the full 1M tokens available for Opus 4.7, just as GitHub Copilot and Claude Code CLI already support.

Zed version and system specs

Zed build: v0.215.0+ (or most recent)
OS: macOS/Linux/Windows (reproducible across environments)
Memory: N/A
Architecture: x86_64/aarch64

Attach Zed log file

Zed.log

Relevant Zed settings

settings.json

"assistant": {
  "default_model": {
    "provider": "anthropic",
    "model": "claude-opus-4-7-1m" 
  },
  "max_tokens": 1000000
}

Relevant Keymap

keymap.json

(for AI issues) Model provider details

• Provider: Anthropic via ZedPro (or API key), Copilot extension
• Model Name: Claude Opus 4.7
• Mode: Assistant Panel, Copilot Bridge
• Workaround: None ; Copilot Pro+ , Business and Enterprise tiers on VS Code and CLI already leverage the full 1M window.

[eiva]

Any estimates on when it can be fixed? It is quite strict limitation...

[nathanchenseanwalter]

Issue is still present. I'm using 1m here.

[mprz1024]

We're impacted by this for Gemini Flash models too. Basically the Zed Agent doesn't seem to let you choose context size?

[ModeSevenIndustrialSolutions]

Would be good to see a fix for this; many of the models support 1M token windows and the new auto-compaction feature needs to be disabled or it has a tendency to kick in prematurely.

[mprz1024]

@ModeSevenIndustrialSolutions , the solution we found was to just install OpenCode, which works well with GitHub Copilot and can be used both in the CLI and within Zed, preview costs, span subagents, access every model, etc.; it has tons of features, and the sessions you start from the Zed UI can be continued in the OpenCode CLI too.

1. You can install the ACP extension for it (Ctrl+Shift+P, ACP registry OR ··· + Add more agents from the agent pane).
2. Go to agents. On top of it, where it says Zed Agent, click and choose OpenCode.
3. Zed will install OpenCode for you, but it may not make the binary available to you. You probably want to fix it with something like

OPENCODE="$(find ~/.local/share/zed/external_agents/registry/opencode -type f -executable -name opencode | sort -n | tail -1)"
[[ $OPENCODE != '' ]] && sudo ln -sfT "$OPENCODE" /usr/local/bin/opencode

4. Once the OpenCode agent opens in Zed, in the ··· at the top of the agent pane, right of OpenCode ∨, choose Reauthenticate. Click the button at the bottom right to log in in a Zed terminal pane. If you do not see the terminal pane, open it manually.
5. Use the arrows to choose GitHub Copilot, then choose GitHub.com (Public)
6. Copy the 8-hex code and go to the URL pointed: https://github.com/login/device
7. Log in with your GitHub Copilot credentials if you have to, enter your 8-hex code, authorise and confirm.

Once you're all set, you'll have access to missing models like Gemini 3.5 Flash (currently the best and most cost effective Western model) and the full 1M context size.